Sunny Lu, associated with VeChain, has shared an extensive analysis arguing that an artificial intelligence market correction is increasingly likely, even if AI ultimately becomes one of the most important technologies of the coming decades. His central argument is that infrastructure investment is moving faster than economically sustainable demand, creating a gap that could trigger financial losses before the technology reaches its full potential.
The analysis compares today’s AI investment cycle with the telecommunications boom of the late 1990s. During the five years after the 1996 Telecommunications Act, carriers invested more than $500 billion in infrastructure, including extensive fiber-optic networks. When the market collapsed in 2000, much of that capacity remained unused, even though the underlying internet technology continued to develop.
AI faces a similar challenge, according to Lu’s argument. Major technology companies are spending enormous sums on computing infrastructure, but investors still need evidence that the resulting capacity can generate sufficient returns. The analysis cites projected hyperscaler capital expenditure approaching $800 billion in 2026, with forecasts rising further in 2027. The key question is whether profitable applications will emerge quickly enough to justify that spending.
Why AI Infrastructure Investment Could Outrun Demand
Lu describes major technology booms as a progression through three stages: infrastructure, applications that improve existing activities, and entirely new products that become possible only after the underlying technology matures. Electricity and the internet followed this pattern, with early infrastructure investment preceding many of the applications that eventually transformed industries.
Today’s AI market, in this framework, has already built much of its foundational infrastructure through GPU clusters and large models. Chatbots represent the first major wave of consumer applications, making existing activities such as searching for information and drafting text faster. However, the analysis argues that these tools have not yet demonstrated the full economic potential needed to support the scale of infrastructure spending.
The next stage could involve AI agents that execute tasks rather than simply provide answers. Instead of generating a logistics recommendation, for example, an agent could contact suppliers, update delivery schedules and issue purchase orders within specified limits. Such systems could compete with the cost of labor and inefficient business processes, potentially opening a much larger market than subscriptions to answer-generating tools.
History also shows why infrastructure booms can become financially destructive. Railway expansion, electricity networks and telecommunications infrastructure were often built ahead of demand. When projected revenues arrived too late to cover borrowing and operating costs, investors suffered even when the underlying infrastructure later proved useful. Lu argues that AI could follow the same pattern: the technology may succeed while some companies financing its expansion fail.
There are important differences between the current AI cycle and the dot-com era. Microsoft, Alphabet, Amazon and Meta have substantial existing businesses that generate cash, potentially making the industry’s core infrastructure investment more resilient than the highly leveraged telecommunications expansion of the late 1990s. However, the analysis warns that smaller companies and more speculative businesses may be more vulnerable if demand disappoints or financing becomes harder to obtain.
NVIDIA, AI Agents and the Race to Generate Returns
NVIDIA occupies a central position in the analysis because it supplies the computing hardware powering much of the AI expansion. Lu compares its position with Cisco’s role during the internet boom. Cisco became one of the world’s most valuable companies around 2000 before collapsing demand from telecommunications customers helped drive its share price down sharply.
NVIDIA’s circumstances differ because its largest customers have considerable financial resources. However, those customers are also developing their own chips. Google’s Tensor Processing Units, Amazon’s Trainium and Inferentia, Microsoft’s Maia and Meta’s MTIA illustrate how major AI buyers are working to reduce dependence on external hardware suppliers. That creates a potential risk for NVIDIA even if overall AI demand remains substantial.
Another challenge is the short economic life of advanced computing hardware. The analysis notes that hyperscalers commonly depreciate servers over roughly five to six years, while some analysts argue that the effective competitive life of leading accelerators can be closer to three years. If newer chips deliver better performance per unit of energy and cost, existing infrastructure could lose economic value before expected returns materialize.
Falling AI inference costs could intensify this pressure while also creating new opportunities. Cheaper intelligence may reduce the value of existing computing capacity, but it can make previously expensive applications economically viable. AI agents that operate continuously, use multiple tools, check their work and complete complex workflows could become practical as the cost of running capable models declines.
Ultimately, Lu’s analysis argues that the outcome depends on whether AI evolves from an answer engine into an action engine. If AI remains largely a tool for generating information, its revenue may struggle to justify the infrastructure bill. If it reliably completes valuable work across business operations, the investment could support a new generation of products and services. A market correction would not necessarily invalidate AI, but it could determine which companies survive and how quickly the technology’s economic benefits emerge.
About The Author
